docs: restructure organization card with 4-tier model catalog and vigyan-pro v4.0.0
Browse files
README.md
CHANGED
|
@@ -1,112 +1,118 @@
|
|
| 1 |
-
---
|
| 2 |
-
title: Experimentlab
|
| 3 |
-
emoji: βοΈ
|
| 4 |
-
colorFrom: indigo
|
| 5 |
-
colorTo: purple
|
| 6 |
-
sdk: static
|
| 7 |
-
pinned: false
|
| 8 |
-
---
|
| 9 |
-
|
| 10 |
<div align="center">
|
| 11 |
|
| 12 |
-
#
|
| 13 |
-
### Sovereign
|
| 14 |
|
| 15 |
-
[](https://discord.com/invite/Jck3tD6mhf)
|
| 20 |
-
[](https://github.com/Vigyan-workstation/vigyan-workstation/releases/tag/v1.0.0)
|
| 21 |
|
| 22 |
-
|
|
|
|
|
|
|
| 23 |
|
| 24 |
</div>
|
| 25 |
|
| 26 |
---
|
| 27 |
|
| 28 |
-
##
|
| 29 |
|
| 30 |
-
**
|
| 31 |
|
| 32 |
-
|
| 33 |
-
* **Zero Cloud Dependence**: 100% offline
|
| 34 |
-
* **Deterministic
|
| 35 |
-
* **
|
| 36 |
|
| 37 |
---
|
| 38 |
|
| 39 |
-
##
|
| 40 |
-
|
| 41 |
-
```mermaid
|
| 42 |
-
graph LR
|
| 43 |
-
A["Raw STEM Corpora (1B+ Tokens, arXiv, Dolma)"] --> B["Vigyan Sovereign Pretraining"]
|
| 44 |
-
B --> C["1. Frontier STEM SLMs (32B / 7B / 2B)"]
|
| 45 |
-
B --> D["2. Advanced Architectures (MoE / 1.58-bit / GRPO)"]
|
| 46 |
-
B --> E["3. Domain Mastery (RTL / Math / Biomed / Vidhi)"]
|
| 47 |
-
C --> F["5. Edge GGUF Runtimes & Desktop Studios"]
|
| 48 |
-
D --> F
|
| 49 |
-
E --> F
|
| 50 |
-
B --> G["4. Foundation Pretraining Datasets"]
|
| 51 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
|
| 53 |
-
|
| 54 |
-
* **
|
| 55 |
-
*
|
| 56 |
-
*
|
| 57 |
-
|
| 58 |
-
#
|
| 59 |
-
* **[`Vigyan-1.5B-4x-MoE`](https://huggingface.co/Experimentlab/Vigyan-1.5B-4x-MoE)** *(1,669+ downloads)*: 4-expert routed Sparse MoE delivering 7B-class STEM reasoning at 1.5B inference latency.
|
| 60 |
-
* **[`Vigyan-3B-4x-MoE`](https://huggingface.co/shreyansh12183/Vigyan-3B-4x-MoE)** *(704+ downloads)*: Advanced Top-1 routing multi-expert model tuned for symbolic logic.
|
| 61 |
-
* **[`vigyan-3b-1.58bit-qat`](https://huggingface.co/shreyansh12183/vigyan-3b-1.58bit-qat)**: Pure 1.58-bit {-1, 0, +1} ternary BitNet model trained via Quantization-Aware Training (QAT). Replaces matrix multiplication with integer additions.
|
| 62 |
-
* **[`vigyan-2b-reasoning-grpo`](https://huggingface.co/shreyansh12183/vigyan-2b-reasoning-grpo)**: DeepSeek-style GRPO test-time reasoning SLM trained on mathematical and logical verification trees.
|
| 63 |
-
* **[`vigyan-olmoe-1b-7b-masterpiece-adapter`](https://huggingface.co/shreyansh12183/vigyan-olmoe-1b-7b-masterpiece-adapter)**: Ai2 OLMoE sparse expert architecture.
|
| 64 |
-
|
| 65 |
-
### 3. π¬ Doctoral & Domain-Specialized Intelligence
|
| 66 |
-
* **[`olmo2-7b-silicon-rtl-eda`](https://huggingface.co/Experimentlab/olmo2-7b-silicon-rtl-eda)**: Hardware silicon EDA specialist for synthesizable IEEE 1364 Verilog RTL, FSMs, and synchronous timing circuits.
|
| 67 |
-
* **[`olmo2-7b-phd-pure-math`](https://huggingface.co/Experimentlab/olmo2-7b-phd-pure-math)**: Specialization in pure mathematics, formal topological proofs, and abstract algebra.
|
| 68 |
-
* **[`olmo2-7b-biomed-chem`](https://huggingface.co/Experimentlab/olmo2-7b-biomed-chem)**: Deep biochemistry, enzyme kinetics, and molecular thermodynamics.
|
| 69 |
-
* **[`olmo2-7b-astro-logic`](https://huggingface.co/Experimentlab/olmo2-7b-astro-logic)**: Orbital mechanics, Hohmann transfers, and astrophysical physics.
|
| 70 |
-
* **[`Vidhi-AI-Instruct`](https://huggingface.co/Experimentlab/Vidhi-AI-Instruct)**: Sovereign Indian legal intelligence aligned with the Bharatiya Nyaya Sanhita (BNS).
|
| 71 |
-
|
| 72 |
-
### 4. π Foundational Pretraining Corpora & Datasets
|
| 73 |
-
* **[`shreyansh-1B-SLM-pretrain-stem-english`](https://huggingface.co/datasets/shreyansh12183/shreyansh-1B-SLM-pretrain-stem-english)**: 1 Billion token foundation pretraining corpus balanced for cold-start STEM intelligence.
|
| 74 |
-
* **[`shreyansh-hinglish-english-stem-500k`](https://huggingface.co/datasets/shreyansh12183/shreyansh-hinglish-english-stem-500k)**: 500,000 sample bilingual Hinglish-English dataset tailored for Indian engineering students.
|
| 75 |
-
* **[`vigyan-ai-research-papers`](https://huggingface.co/datasets/Experimentlab/vigyan-ai-research-papers)**: Curated corpus of scientific literature and verified technical preprints.
|
| 76 |
-
|
| 77 |
-
### 5. β‘ Edge GGUF Runtimes & Workstation Integration
|
| 78 |
-
* **[`Vigyan-Models-GGUF`](https://huggingface.co/Experimentlab/Vigyan-Models-GGUF)**: Quantized GGUF binaries (`Q4_K_M`) engineered for local execution via llama.cpp and Ollama.
|
| 79 |
-
* **[`vigyan-7b-moe-gguf`](https://huggingface.co/shreyansh12183/vigyan-7b-moe-gguf)**: Edge quantized MoE binary.
|
| 80 |
-
* **[Vigyan Workstation](https://github.com/Vigyan-workstation/vigyan-workstation)**: Open-source desktop engineering studio pairing these weights with 3D robotics kinematics, WaveDrom waveforms, and KiCad PCB netlists.
|
| 81 |
|
| 82 |
---
|
| 83 |
|
| 84 |
-
##
|
| 85 |
|
| 86 |
-
All
|
| 87 |
|
| 88 |
-
* **Academic, Educational & Personal Research Use
|
| 89 |
-
* **
|
| 90 |
-
* **
|
| 91 |
-
* **Commercial Enterprise Licensing**: For commercial licensing, enterprise deployment, or
|
| 92 |
|
| 93 |
---
|
| 94 |
|
| 95 |
-
##
|
| 96 |
|
| 97 |
```bibtex
|
| 98 |
@article{singh2026vigyan,
|
| 99 |
title={Vigyan Sovereign Trilogy: Empirical Scaling Laws and Deterministic Symbolic-AST Evaluation of Edge Small Language Models},
|
| 100 |
author={Singh, Shreyansh},
|
| 101 |
-
institution={
|
| 102 |
-
journal={Vigyan AI Research Publications},
|
| 103 |
year={2026},
|
| 104 |
-
url={https://github.com/
|
| 105 |
}
|
| 106 |
```
|
| 107 |
|
| 108 |
---
|
| 109 |
|
| 110 |
<div align="center">
|
| 111 |
-
<sub>
|
| 112 |
</div>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
<div align="center">
|
| 2 |
|
| 3 |
+
# ExperimentLab AI
|
| 4 |
+
### Sovereign Physical Engineering Intelligence & Foundation Small Language Models
|
| 5 |
|
| 6 |
+
[](https://github.com/Vigyan-workstation)
|
| 7 |
+
[](https://github.com/Vigyan-workstation/vigyan-pro)
|
| 8 |
+
[](https://discord.com/invite/Jck3tD6mhf)
|
| 9 |
+
[](https://github.com/Vigyan-workstation/vigyan-pro/blob/main/LICENSE)
|
|
|
|
|
|
|
| 10 |
|
| 11 |
+
<p align="center">
|
| 12 |
+
<b>Pioneering sovereign edge intelligence, air-gapped engineering desktop runtimes, Sparse Mixture of Experts (MoE), and deterministic Computer Algebra System (SymPy CAS) verification.</b>
|
| 13 |
+
</p>
|
| 14 |
|
| 15 |
</div>
|
| 16 |
|
| 17 |
---
|
| 18 |
|
| 19 |
+
## Mission & Architecture
|
| 20 |
|
| 21 |
+
**ExperimentLab AI** designs and trains foundation Small Language Models (SLMs) and sovereign desktop workstations specifically for **physical engineers, mathematicians, roboticists, and silicon designers**.
|
| 22 |
|
| 23 |
+
Traditional cloud language models hallucinate mathematical proofs, leak sensitive IP via telemetry, and require costly continuous token fees. ExperimentLab eliminates these barriers through a dual-engine architecture:
|
| 24 |
+
* **Zero Cloud Dependence**: 100% offline edge execution on local hardware with zero network egress.
|
| 25 |
+
* **Deterministic CAS Truth**: Neural natural-language reasoning verified through in-process SymPy Computer Algebra System AST kernels.
|
| 26 |
+
* **Edge-Optimized Architectures**: Sparse Mixture of Experts (MoE) and ternary 1.58-bit {-1, 0, +1} BitNet quantization running natively on standard consumer CPUs without requiring discrete GPUs.
|
| 27 |
|
| 28 |
---
|
| 29 |
|
| 30 |
+
## 4-Tier Public Model Catalog
|
| 31 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
```
|
| 33 |
+
βββββββββββββββββββββββββββββββββββββββββββββββββοΏ½οΏ½ββββββββββββββββββββββββ
|
| 34 |
+
β ExperimentLab Foundation Model & Runtime Architecture β
|
| 35 |
+
β β
|
| 36 |
+
β [Tier 1] Edge GGUF Runtimes (Vigyan-Models-GGUF for vigyan CLI) β
|
| 37 |
+
β [Tier 2] Sparse Mixture of Experts & Ternary Quantization (MoE, QAT) β
|
| 38 |
+
β [Tier 3] Dense Foundation STEM SLMs (32B Titan, 7B, 2B Edge) β
|
| 39 |
+
β [Tier 4] Doctoral Domain Checkpoints (Silicon RTL, Math, BioMed) β
|
| 40 |
+
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 41 |
+
```
|
| 42 |
+
|
| 43 |
+
### Tier 1: Production Edge GGUF Runtimes
|
| 44 |
+
Optimized quantized weights engineered for the **[Vigyan Pro Desktop Workstation](https://github.com/Vigyan-workstation/vigyan-pro)** and CLI model management:
|
| 45 |
+
* **[`Experimentlab/Vigyan-Models-GGUF`](https://huggingface.co/Experimentlab/Vigyan-Models-GGUF)**: Verified `Q4_K_M` and `IQ3_M` quantized binaries for `vigyan-2b-reasoning`, `vigyan-2b-stem`, and `vigyan-7b-stem`. Executable on standard laptops with less than 4 GB RAM.
|
| 46 |
+
|
| 47 |
+
Pull directly via the Vigyan CLI:
|
| 48 |
+
```bash
|
| 49 |
+
vigyan model pull vigyan-2b-reasoning-q4_k_m.gguf
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
### Tier 2: Flagship Sparse Mixture of Experts (MoE) & Ternary Quantization
|
| 53 |
+
* **[`Experimentlab/Vigyan-1.5B-4x-MoE`](https://huggingface.co/Experimentlab/Vigyan-1.5B-4x-MoE)**: 4-expert routed Sparse MoE providing 7B-class STEM reasoning at 1.5B active parameter inference speed.
|
| 54 |
+
* **[`shreyansh12183/Vigyan-3B-4x-MoE`](https://huggingface.co/shreyansh12183/Vigyan-3B-4x-MoE)**: Top-1 routing multi-expert model calibrated for symbolic deduction.
|
| 55 |
+
* **[`shreyansh12183/vigyan-3b-1.58bit-qat`](https://huggingface.co/shreyansh12183/vigyan-3b-1.58bit-qat)**: Pure 1.58-bit {-1, 0, +1} ternary BitNet model trained via Quantization-Aware Training.
|
| 56 |
+
* **[`shreyansh12183/vigyan-2b-reasoning-grpo`](https://huggingface.co/shreyansh12183/vigyan-2b-reasoning-grpo)**: GRPO test-time reasoning SLM trained on mathematical and logical verification trees.
|
| 57 |
+
|
| 58 |
+
### Tier 3: Foundation STEM Dense SLMs
|
| 59 |
+
* **[`Experimentlab/Vigyan-AI-32B`](https://huggingface.co/Experimentlab/Vigyan-AI-32B)** & **[`Experimentlab/Vigyan-AI-32B-Titan-v1`](https://huggingface.co/Experimentlab/Vigyan-AI-32B-Titan-v1)**: Flagship 32B models for deep multi-step scientific deduction and formal derivations.
|
| 60 |
+
* **[`Experimentlab/Vigyan-7B-STEM-Instruct-v1`](https://huggingface.co/Experimentlab/Vigyan-7B-STEM-Instruct-v1)** & **[`Experimentlab/Shreyansh-STEM-AI-7B-Final`](https://huggingface.co/Experimentlab/Shreyansh-STEM-AI-7B-Final)**: 7B dense scientific reasoning models.
|
| 61 |
+
* **[`Experimentlab/Shreyansh-STEM-AI-2B-v3`](https://huggingface.co/Experimentlab/Shreyansh-STEM-AI-2B-v3)** & **[`Experimentlab/Shreyansh-STEM-AI-2B-v3-Distilled`](https://huggingface.co/Experimentlab/Shreyansh-STEM-AI-2B-v3-Distilled)**: Compact 2B parameter edge SLMs for battery-powered operation.
|
| 62 |
+
* **[`Experimentlab/Vigyan-2B-STEM-Instruct-v1`](https://huggingface.co/Experimentlab/Vigyan-2B-STEM-Instruct-v1)**: Instruction-tuned STEM canary model.
|
| 63 |
+
|
| 64 |
+
### Tier 4: Doctoral & Domain-Specialized Intelligence
|
| 65 |
+
* **[`Experimentlab/olmo2-7b-silicon-rtl-eda`](https://huggingface.co/Experimentlab/olmo2-7b-silicon-rtl-eda)**: Hardware silicon specialist for synthesizable IEEE 1364 Verilog RTL, finite state machines, and synchronous timing circuits.
|
| 66 |
+
* **[`Experimentlab/olmo2-7b-phd-pure-math`](https://huggingface.co/Experimentlab/olmo2-7b-phd-pure-math)**: Pure mathematics, formal topology, real analysis, and abstract algebra.
|
| 67 |
+
* **[`Experimentlab/olmo2-7b-biomed-chem`](https://huggingface.co/Experimentlab/olmo2-7b-biomed-chem)**: Deep biochemistry, enzyme kinetics, and molecular thermodynamics.
|
| 68 |
+
* **[`Experimentlab/olmo2-7b-astro-logic`](https://huggingface.co/Experimentlab/olmo2-7b-astro-logic)**: Orbital mechanics, Hohmann transfers, and astrophysical dynamics.
|
| 69 |
+
* **[`Experimentlab/Vidhi-AI-Instruct`](https://huggingface.co/Experimentlab/Vidhi-AI-Instruct)**: Sovereign Indian legal intelligence aligned with the Bharatiya Nyaya Sanhita (BNS).
|
| 70 |
+
|
| 71 |
+
### Open Research Corpora & Datasets
|
| 72 |
+
* **[`Experimentlab/vigyan-ai-research-papers`](https://huggingface.co/datasets/Experimentlab/vigyan-ai-research-papers)**: Scientific literature and verified technical preprints.
|
| 73 |
+
* **[`shreyansh12183/shreyansh-1B-SLM-pretrain-stem-english`](https://huggingface.co/datasets/shreyansh12183/shreyansh-1B-SLM-pretrain-stem-english)**: 1 Billion token foundation pretraining corpus balanced for cold-start STEM intelligence.
|
| 74 |
+
* **[`shreyansh12183/shreyansh-hinglish-english-stem-500k`](https://huggingface.co/datasets/shreyansh12183/shreyansh-hinglish-english-stem-500k)**: 500,000 sample bilingual Hinglish-English dataset tailored for Indian engineering students.
|
| 75 |
+
|
| 76 |
+
---
|
| 77 |
+
|
| 78 |
+
## Community & Support
|
| 79 |
+
|
| 80 |
+
Join physical engineers, researchers, and autonomous agent builders on our official Discord server:
|
| 81 |
|
| 82 |
+
* **Discord Server**: **[Experimentlab Community](https://discord.com/invite/Jck3tD6mhf)** (65+ Members)
|
| 83 |
+
* **Key Channels**:
|
| 84 |
+
* `#announcements`: New model weights, benchmark scorecards, and research releases.
|
| 85 |
+
* `#workstation`: Vigyan Pro desktop installation, CLI model pulling, and setup support.
|
| 86 |
+
* `#research`: Model pretraining, MoE upcycling, and symbolic CAS evaluation.
|
| 87 |
+
* `#model-evals`: GSM8K, MATH, and Verilog synthesis test results.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
|
| 89 |
---
|
| 90 |
|
| 91 |
+
## Legal, Intellectual Property & Non-Commercial Licensing
|
| 92 |
|
| 93 |
+
All models, quantized weights, pretraining datasets, research papers, and software released by ExperimentLab AI and Vigyan Workstation are strictly licensed under the **PolyForm Noncommercial License 1.0.0**:
|
| 94 |
|
| 95 |
+
* **Academic, Educational & Personal Research Use**: Freely permitted for students, academic researchers, university coursework, self-hosting, and civilian scientific inquiry.
|
| 96 |
+
* **AI Model Distillation Restriction (Section 8)**: You may not use any model outputs, weights, datasets, or documentation to train, fine-tune, distill, or evaluate competing artificial intelligence or foundation models without prior written enterprise authorization.
|
| 97 |
+
* **Trademark Protection (Section 9)**: No trademark, service mark, or brand rights in "ExperimentLab AI" or "Vigyan Pro" are granted.
|
| 98 |
+
* **Commercial Enterprise Licensing**: For commercial licensing, enterprise deployment, or OEM hardware distribution, contact **licensing@experimentlab.in**.
|
| 99 |
|
| 100 |
---
|
| 101 |
|
| 102 |
+
## Academic Attribution & Sovereign Citation
|
| 103 |
|
| 104 |
```bibtex
|
| 105 |
@article{singh2026vigyan,
|
| 106 |
title={Vigyan Sovereign Trilogy: Empirical Scaling Laws and Deterministic Symbolic-AST Evaluation of Edge Small Language Models},
|
| 107 |
author={Singh, Shreyansh},
|
| 108 |
+
institution={ExperimentLab AI Open Research},
|
|
|
|
| 109 |
year={2026},
|
| 110 |
+
url={https://github.com/Vigyan-workstation/vigyan-pro}
|
| 111 |
}
|
| 112 |
```
|
| 113 |
|
| 114 |
---
|
| 115 |
|
| 116 |
<div align="center">
|
| 117 |
+
<sub>Developed by <a href="https://experimentlab.in">ExperimentLab AI</a>. Built for physical engineers who value sovereignty and mathematical precision.</sub>
|
| 118 |
</div>
|